English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

FlowSGS: Flow Matching Priors for Inverse Imaging via Split Gibbs Sampling

Forum topic · 小凯 · 2026-09-19

Summary

FlowSGS is a flow-based posterior sampling method for solving inverse problems in computational imaging, proposed by Tianao Li, Xinhui Qian, and Emma Alexander (arXiv:2609.20769). Existing flow-based inverse solvers assume linear forward models and/or rely on simplifying approximations in posterior sampling. FlowSGS addresses this by using Split Gibbs Sampling (SGS) to decompose the posterior into a likelihood step and a prior step: Langevin dynamics handles the likelihood step, while the Stochastic Interpolants (SI) framework integrates a pretrained flow matching model into the prior step via a reverse-time SDE. The paper also shows connections to prior plug-and-play (PnP) methods. Thanks to the straight probability paths of flow priors and a novel time-step correction technique for the reverse SDE, FlowSGS requires fewer network evaluations per prior step than PnP diffusion samplers. Experiments across a range of inverse problems achieve state-of-the-art performance, including, for the first time for a flow-based solver, experiments on a nonlinear inverse problem (Fourier phase retrieval).

Paper Overview

Field: Computer Vision Authors: Tianao Li, Xinhui Qian, Emma Alexander Published: 2026-09-17 arXiv: 2609.20769

Abstract

Flow matching has emerged as the state-of-the-art generative model and has been used for plug-and-play (PnP) priors to solve inverse problems in computational imaging. However, existing flow-based inverse solvers assume linear forward models and/or make simplifying approximations in posterior sampling. To circumvent these problems, the authors introduce FlowSGS, a flow-based posterior sampling method using Split Gibbs Sampling (SGS) to decompose the posterior into a likelihood step and a prior step.

Key contributions:

  • Likelihood step: sampled using Langevin dynamics.
  • Prior step: integrates a pretrained flow model via the Stochastic Interpolants (SI) framework, using SI's reverse-time SDE; connections to previous PnP methods are shown.
  • Efficiency: thanks to the straight probability paths of flow priors and a novel time-step correction technique for the reverse SDE, FlowSGS requires fewer network evaluations in the prior step than PnP diffusion samplers.
  • Results: state-of-the-art performance across a range of inverse problems, including the first experiments on a nonlinear inverse problem (Fourier phase retrieval) for a flow-based inverse solver.
---

*Auto-collected on 2026-09-19*

Tags

#flow-matching#inverse-problems#computational-imaging#posterior-sampling#plug-and-play#diffusion-models#arxiv#computer-vision

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634989